Accurate QoT estimation for the optimized design of optical transport network based on advanced deep learning model

Increasing traffic demand in the backbone optical network urges the initiation of new solutions that allow an increment in the transmission/network capacity without a considerable rise in service costs. To attain this objective, simpler and quicker network reconfiguration processes is required. A significant issue intrinsically correlated with network reconfiguration is the quality of transmission (QoT) (BER, SNR, Q-factor, chromatic dispersion and polarization mode dispersion) estimation. Predicting the QoT of already established light-paths enable network operators to respond proactively to performance deterioration, which helps in enhancing or sustaining quality of service (QoS). Recently, several deep learning-based approaches have been presented to estimate the QoT of an optical network. This paper proposes an advanced deep learning model comprising a long short-term memory (LSTM) network along with three fully connected layers that leverage the transmission quality metrics collected from the network's control plane to predict QoT of light-path. The three additional fully connected layers help the model learn the more complex patterns in the data that a basic LSTM cannot learn on its own and improve the model's performance by three orders of magnitude. Moreover, the proposed model doesn't require a large data-set to provide high accuracy results, which decreases the control plane's load. Our model was tested on a data-set that contains data extracted from Microsoft's optical backbone network in North America. We compare the performance of the proposed model with reference approaches viz; SVM, MLP, NN, and LSTM to show if such models could be considered for predicting the estimated received Q-factor and chromatic dispersion of established light-paths. Results exhibit that the proposed model achieves the best performance across multiple channels of this data-set and even in a case when the size of the training data-set is small. The proposed model provides high prediction accuracy, which is perfect for the next-generation agile optical network.